PoMeS: Profit-Maximizing Sensor Selection for Crowd-Sensed ... · PoMeS: Profit-Maximizing Sensor...

PoMeS: Profit-Maximizing Sensor Selection for Crowd-Sensed Spectrum Discovery

Suzan Bayhan TU Berlin, Germany

https://suzanbayhan.github.io/

Joint work with: Gürkan Gür (ZHAW, Switzerland), Anatolij Zubow (TU Berlin, Germany)

https://suzanbayhan.github.io/

/25PoMeS

High variation in a cellular network’s load

Xu, F., Li, Y., Wang, H., Zhang, P., Jin, D.: Understanding mobile traffic patterns of large scale cellular towers in urban environment. IEEE/ACM TON, 2017. !2

/25PoMeS



~8x

/25PoMeS



~8x

Need for over-provisioning

/25PoMeS



Capacity expansion via secondary spectrum rather than costly capacity over-provisioning

~8x

Need for over-provisioning

/25PoMeS

Crowd-sourcing based spectrum-discovery • Rather than deploying its own infrastructure, the MNO

launches crowd-sensing campaign

!3

Crowd-sourcing Spectrum Sensing

Spectrum sensor selected for sensing

/25PoMeS



!4

Sensor selection

Measurements


Spectrum fusion


/25PoMeS



!4

Sensor selection

Measurements


Spectrum fusion


APPs

/25PoMeS

Applications of crowd-sourcing based spectrum sensing

• Spectrum monitoring for better policy making

• Spectrum patrolling for detecting spectrum misuse

– Chakraborty et al. Spectrum patrolling with crowdsourced spectrum sensors, IEEE INFOCOM 2018

• Radio Environment Map generation and spectrum queries

– Chakraborty et al. Specsense: Crowd-sensing for efficient querying of spectrum occupancy, IEEE INFOCOM 2017

– Ying et al. Pricing mechanism for quality-based radio mapping via crowdsourcing, IEEE GLOBECOM 2016

!5

/25PoMeS

Applications of crowd-sourcing based spectrum sensing

• Spectrum monitoring for better policy making

• Spectrum patrolling for detecting spectrum misuse

– Chakraborty et al. Spectrum patrolling with crowdsourced spectrum sensors, IEEE INFOCOM 2018

• Radio Environment Map generation and spectrum queries

– Chakraborty et al. Specsense: Crowd-sensing for efficient querying of spectrum occupancy, IEEE INFOCOM 2017

– Ying et al. Pricing mechanism for quality-based radio mapping via crowdsourcing, IEEE GLOBECOM 2016

!5

PoMeS: crowd-sourcing based spectrum discovery for MNO capacity expansion

/25PoMeS

PoMeS: profit-maximizing sensor selection

!6

Cell loadstatistics

PU activitystatistics

SensorselectionProfit model

MNO network management

Hotspot cell withhigh traffic load

Coldspot cell with lowtraffic load

MNO

customers

Spectrum sensor not

selected for sensing

Spectrum sensor


/25PoMeS

PoMeS: profit-maximizing sensor selection• How many sensors to use for spectrum

discovery?

– Monetary cost of spectrum sensing

– A limited budget for crowd-sensors

– Expected traffic in each cell

• Hot spot cells vs cold spot cells

• Varying expected PU traffic

– Required sensing accuracies asserted by the regulatory bodies

!6

Cell loadstatistics

PU activitystatistics

SensorselectionProfit model

MNO network management

Hotspot cell withhigh traffic load

Coldspot cell with lowtraffic load

MNO

customers

Spectrum sensor not


Spectrum sensor


/25PoMeS

Goal: maximize the profit while meeting the regulatory requirements

!7

• Regulations: might be overly-conservative resulting in wasteful sensing by the sensors

– High PU detection accuracy (>0.90)

– Low false alarm probability (<0.10)

• Oblivious to the PU traffic or secondary network’s traffic

/25PoMeS

Goal: maximize the profit while meeting the regulatory requirements

!7

• Regulations: might be overly-conservative resulting in wasteful sensing by the sensors

– High PU detection accuracy (>0.90)

– Low false alarm probability (<0.10)

• Oblivious to the PU traffic or secondary network’s traffic

PoMeS: different accuracy at each cell, but monetary penalty if the required accuracy not met

/25PoMeS

Spectrum-sensing model

!8

• PU statistics are available at the MNO

• Sensor accuracies are identical and Pd, Pf known by the MNO

• Sensors’ sensing price is identical

• Majority decision combining

• Sensing period, reporting period

Sensing period

TDMA reporting

period

Transmission period in case of idle spectrum

T

MAJORITY LOGIC

/25PoMeS

MNO’s net profit

!9

Expenses for crowd-spectrum sensing

Penalty paid if spectrum sensing

accuracy is lower than required

from its customers served through the

discovered spectrum

Sensing cost Collision cost Income

/25PoMeS

MNO’s net profit

!9





discovered spectrum

How many sensors are selected? Price of each sensor


/25PoMeS

MNO’s net profit

!10





discovered spectrum

Achieved sensing accuracy Required accuracy

Penalty policy


/25PoMeS

MNO’s net profit

!11





discovered spectrum

Demand in each cell Discovered spectrum in each cell

Price of each served request


/25PoMeS

Utility of spectrum sensing with m sensors

• Utility U(m): expected discovered and useable spectrum if m sensors sense the spectrum

!12

Probability that PU channel is idle

PU channel’s bandwidth (Hz)

Sensing efficiency (after overheads of

sensing and reporting)

Probability of false alarms in sensing

𝒰(m) = poB ( T − Ts − mTr

T )(1 − Qf(m))

/25PoMeS

How many requests can be served with this discovered spectrum?

!13

Number of requests in this cell-i

Required minimum resources per request

Spectral efficiency of the MNO

Rmaxi = min(ri,

𝒰iκcmin

). requests/sec

Discovered spectrum

/25PoMeS

Income of cell-i

• Each served request translates into some monetary gain

!14

Service cost paid for each served request

Rmaxi = min(ri,

𝒰iκcmin

). requests/sec

Π+i = μRmax

i

/25PoMeS

Expenses of cell-i : sensing cost

!15

Unit sensing price

Frequency of sensing

Number of sensors

Π−i = Niμsβs

Sensing cost Collision penalty cost

ΔQd,i = max(0,Q*d − Qd,i)

μcΔQd,iRimax

Penalty price

Difference between the desired and achieved sensing

accuracy

/25PoMeS

Optimal sensor selection problem

!16

Available budget for sensors

maxNi

∑Ai∈𝒜

Rmaxi μ − Niμsβs − μcRmax

i max(0,ΔQd,i − Q*)

∑Ai∈𝒜

μsβsNi ⩽ ℬ

Ni ⩽ ⌊T − Ts

Tr⌋

Ni ⩾ 0

/25PoMeS

Optimal sensor selection problem

!16

Available budget for sensors

maxNi

∑Ai∈𝒜

Rmaxi μ − Niμsβs − μcRmax

i max(0,ΔQd,i − Q*)

∑Ai∈𝒜

μsβsNi ⩽ ℬ

Ni ⩽ ⌊T − Ts

Tr⌋

Ni ⩾ 0Coupling constraint. NP-hard! De-couple via allocating the budget first.

(i) budget allocation problem (ii) exhaustive search in each cell

/25PoMeS

Polynomial complexity:

EQ, PROP:

INGA:

• Equal budget per cell (EQ):

• # of sensors upper-bounded by:

• PROP: Budget proportional to the serving capacity of the cell

• Incremental gain based greedy assignment (INGA)

• Baselines:

• satisfying (Q∗d,Q∗

f) required by the regulatory body (REG) with EQ or PROP budget allocation

udget allocation for K cells

!17

Nmax = min(⌊T − Ts

Tr⌋, ⌊

ℬKμsβs

⌋)

ℬ

ℬi =Rmax

i ℬ∑Ai∈𝒜 Rmax

i

𝒪(KNmax)

𝒪(KN log(N))

/25PoMeS

Simulation-based performance analysis of PoMeS

!18

• Impact of increasing budget

• Impact of cell traffic load

• Impact of hot-spots (cell-load variation)

/25PoMeS

Parameters

• K = 2000 cell sites

• PU activity = [0.2, 0.8]

• μs = 1, μ = 1, μc = 5,

• κ = 10 bps/Hz, (Pd,Pf)=(0.8, 0.1), and (Q∗d,Q∗f )= (0.98, 0.05)

• Randomly σ of the cells as hotspots

• Rσ fraction of the requests from hotspots

• Coldspot traffic: (1-Rσ) fraction of the requests

!19

/25PoMeS

Impact of budget: B = [1-10] sensors/cell

!20

/25PoMeS


!20

~2x

/25PoMeS


!20

~2x

~0x

/25PoMeS


!21

~2x

~0x

/25PoMeS


!21

~2x

~0x

Saturation in profit due to diminishing returns: deploying more sensors only increases the capacity

marginally

/25PoMeS


!22

/25PoMeS


!22

Under low budget, sacrifice from sensing accuracy

/25PoMeS

Which cells enjoy the capacity expansion?

!23

/25PoMeS

Which cells enjoy the capacity expansion?

!23

Low budget: capacity expansion over all cells with our heuristics

17% of the cell sites vs 65-100%

/25PoMeS

Impact of cell-load

!24

/25PoMeS

Impact of cell-load

!24

- INGA > PROP or EQ by about 5%

- REG-EQ over-performs REG-PROP for about (5-15%) depending on the setting

- Lower sensing accuracy only under low load

/25PoMeS

Impact of hot-spots

!25

Jain’s fairness index in terms of cell load

/25PoMeS

Impact of hot-spots

!25

• Under a more uniform traffic load, profit is higher

• The relative performance of our schemes exhibit the same trend

Jain’s fairness index in terms of cell load

/25PoMeS

Take-aways• Problem:

• Capacity over-provisioning results in a high cost at an MNO • PoMeS:

• Capacity expansion via opportunistic spectrum access • Crowd-sourced spectrum sensing • Select sensors considering MNO’s net profit

• Load of each cell, PU spectrum activity, required spectrum sensing accuracy, each sensor’s cost and accuracy

• Key results • Lower sensing accuracy only when the network load is low and

budget for spectrum sensing payment is limited • Distributing the budget equally for regulation-confirming schemes

results in higher profit • Future work: heterogenous sensors in terms of accuracy and cost

!26

/25PoMeS

Take-aways• Problem:

• Capacity over-provisioning results in a high cost at an MNO • PoMeS:

• Capacity expansion via opportunistic spectrum access • Crowd-sourced spectrum sensing • Select sensors considering MNO’s net profit

• Load of each cell, PU spectrum activity, required spectrum sensing accuracy, each sensor’s cost and accuracy

• Key results • Lower sensing accuracy only when the network load is low and

budget for spectrum sensing payment is limited • Distributing the budget equally for regulation-confirming schemes

results in higher profit • Future work: heterogenous sensors in terms of accuracy and cost

!26

Thank you

PoMeS: Profit-Maximizing Sensor Selection for Crowd-Sensed ... · PoMeS: Profit-Maximizing Sensor...

Documents

Transcript of PoMeS: Profit-Maximizing Sensor Selection for Crowd-Sensed ... · PoMeS: Profit-Maximizing Sensor...